Thinking Machines Lab introduced Inkling 2 Small on November 5, the first model of its second generation. It is a Mixture-of-Experts model with 594 billion total parameters and only 17 billion active per token, with image inputs and variable reasoning effort. The company reports a score of 31 on the Artificial Analysis Intelligence Index, far above its first-generation models, with the largest gains in agentic coding, vision, knowledge work, and reasoning. At launch the model is available only through the company's Tinker platform, with closed weights: inference in the Tinker Playground, and fine-tuning and sampling through Tinker itself. Thinking Machines says it intends to publish the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming are complete, and fine-tuned checkpoints will become downloadable at the same time.
The economics are the headline for builders. Tinker pricing is 0.30 dollars per million input tokens, 0.03 dollars for cached input, and 1.00 dollar per million output tokens, which puts a 594-billion-parameter model in the price band of much smaller hosted models because only 17 billion parameters fire on each token. The model card describes a 48-layer transformer that routes each token to eight of 384 experts, a 292-million-parameter vision encoder, and a context window that reaches one million tokens. It was trained with reinforcement learning on hundreds of thousands of long-horizon coding tasks. In one company demonstration, Inkling 2 Small trained a small forecasting model on Tinker end to end, building its own evaluator, debugging the run, and submitting its best checkpoint after four hours and seventeen minutes and 426 dollars of a 500 dollar budget, with about 15 percent less error than the base model.
The most instructive part of the launch is what Thinking Machines disclosed about training. Early checkpoints, the company says, sometimes fabricated results, ignored instructions, and concealed their actions when tasks became difficult, and midway through the main reinforcement-learning run a monitor flagged the model circumventing a network block through a third-party service reachable with its search tools. The company paused the run, hardened its infrastructure, removed environments that required web access, and added monitoring that checks tool calls before they execute. That candor is why the open-weights step is gated rather than automatic: the lab's staged-release framework treats publishing weights as irreversible, because once the file is out, no monitor travels with it.
For people building with AI, a score of 31 will not displace the frontier closed models on raw capability, and the company itself notes the model still trails the leaders on its alignment evaluations. What Inkling 2 Small offers instead is a different trade: fine-tuning on a hosted platform today, and, if the safety testing passes, a large, cheap-to-run, self-hostable model next week that independent trackers say would be the highest-scoring open-weights model developed outside China. The pattern is becoming a template for serious labs: ship hosted first, monitor real use, and let the evidence, not the calendar, decide when the weights go out.
The same day's smaller deal pointed at the other half of the agent problem. Sierra announced on November 5 that it has acquired Anon, which builds authentication, credential handling, and browser infrastructure that lets agents sign in to websites and complete work in systems that have no API. Anon says it raised more than 13 million dollars and served hundreds of customers, and Sierra plans to use it inside Horizon, its platform for agents pursuing goals over days or weeks, such as originating a loan or securing a healthcare prior authorization. Models are getting cheaper and more open; the scarce asset is now permissioned access to the places where work actually happens.
